Control-oriented modelling for model predictive control of water reservoir systems

When dealing with systems subject to unpredictable external disturbances, choosing the prediction model to incorporate into a Model Predictive Controller (MPC) can be a challenge. We demonstrate that for a fixed model class, the identification through simultaneous fitting of historical data and minimisation of a backtesting control cost allows one to improve performance compared to using all data exclusively for historical fitting. We show that the best prediction model from a control perspective is not necessarily the one minimising the identification cost. Here, system identification takes place online, thanks to Bayesian Optimisation, which iteratively provides the best prediction model, bringing the controller close to the ideal performance. The approach is demonstrated on the control of Lake Como, Italy, a regulated lake operated for irrigation and flood control. The methodology is promising particularly in dry periods, where the impact of predictions is fundamental in hedging irrigation releases to the agricultural district.